[X86][Clang] Add AVX512 kunpck intrinsics to be used in constexp

clang/lib/AST/ExprConstant.cpp

@@ -16287,6 +16287,42 @@ bool IntExprEvaluator::VisitBuiltinCallExpr(const CallExpr *E,
     return Success((A | B) == 0, E);
   }
 
+  case clang::X86::BI__builtin_ia32_kunpckhi: {
+    APSInt A, B;
+    if (!EvaluateInteger(E->getArg(0), A, Info) ||
+        !EvaluateInteger(E->getArg(1), B, Info))
+      return false;
+
+    // Extract lower 8 bits of each operand and concatenate
+    // Result = (A[7:0] << 8) | B[7:0]
+    APSInt Result = ((A & 0xFF) << 8) | (B & 0xFF);
+    return Success(Result, E);
+  }
+
+  case clang::X86::BI__builtin_ia32_kunpckdi: {
+    APSInt A, B;
+    if (!EvaluateInteger(E->getArg(0), A, Info) ||
+        !EvaluateInteger(E->getArg(1), B, Info))
+      return false;
+
+    // Extract lower 32 bits of each operand and concatenate
+    // Result = (A[31:0] << 32) | B[31:0]
+    APSInt Result = ((A & 0xFFFFFFFFULL) << 32) | (B & 0xFFFFFFFFULL);
+    return Success(Result, E);
+  }
+
+  case clang::X86::BI__builtin_ia32_kunpcksi: {
+    APSInt A, B;
+    if (!EvaluateInteger(E->getArg(0), A, Info) ||
+        !EvaluateInteger(E->getArg(1), B, Info))
+      return false;
+
+    // Extract lower 16 bits of each operand and concatenate
+    // Result = (A[15:0] << 16) | B[15:0]
+    APSInt Result = ((A & 0xFFFF) << 16) | (B & 0xFFFF);
+    return Success(Result, E);
+  }
+
   case clang::X86::BI__builtin_ia32_lzcnt_u16:
   case clang::X86::BI__builtin_ia32_lzcnt_u32:
   case clang::X86::BI__builtin_ia32_lzcnt_u64: {
@@ -16413,6 +16449,21 @@ bool IntExprEvaluator::VisitBuiltinCallExpr(const CallExpr *E,
     return Success(APValue(Result), E);
   }
 
+  case X86::BI__builtin_ia32_kunpckhi:
+  case X86::BI__builtin_ia32_kunpcksi:
+  case X86::BI__builtin_ia32_kunpckdi: {
+    return HandleMaskBinOp([](const APSInt &LHS, const APSInt &RHS) {
+      // Unpack: concatenate lower half of RHS with lower half of LHS
+      unsigned HalfBits = LHS.getBitWidth() / 2;
+      APSInt Mask = APSInt::getMaxValue(LHS.getBitWidth(), LHS.isUnsigned());
+      Mask = Mask.trunc(HalfBits).zext(LHS.getBitWidth());
+
+      APSInt LowerLHS = LHS & Mask;
+      APSInt LowerRHS = RHS & Mask;
+      return LowerRHS | (LowerLHS << HalfBits);
+    });
+  }
+
   case X86::BI__builtin_ia32_kaddqi:
   case X86::BI__builtin_ia32_kaddhi:
   case X86::BI__builtin_ia32_kaddsi:

clang/lib/Headers/avx512bwintrin.h

@@ -1606,15 +1606,14 @@ _mm512_maskz_set1_epi8(__mmask64 __M, char __A) {
                                               (__v64qi) _mm512_setzero_si512());
 }
 
-static __inline__ __mmask64 __DEFAULT_FN_ATTRS _mm512_kunpackd(__mmask64 __A,
-                                                               __mmask64 __B) {
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS_CONSTEXPR
+_mm512_kunpackd(__mmask64 __A, __mmask64 __B) {
   return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A,
                 (__mmask64) __B);
 }
 
-static __inline__ __mmask32 __DEFAULT_FN_ATTRS
-_mm512_kunpackw (__mmask32 __A, __mmask32 __B)
-{
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS_CONSTEXPR
+_mm512_kunpackw(__mmask32 __A, __mmask32 __B) {
   return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A,
                 (__mmask32) __B);
 }

clang/lib/Headers/avx512fintrin.h

@@ -40,6 +40,8 @@ typedef long long __m512i_u __attribute__((__vector_size__(64), __aligned__(1)))
 
 typedef unsigned char __mmask8;
 typedef unsigned short __mmask16;
+typedef unsigned int __mmask32;
+typedef unsigned long long __mmask64;
 
 /* Rounding mode macros.  */
 #define _MM_FROUND_TO_NEAREST_INT   0x00
@@ -8094,12 +8096,21 @@ _kortest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C) {
   return (unsigned char)__builtin_ia32_kortestzhi(__A, __B);
 }
 
-static __inline__ __mmask16 __DEFAULT_FN_ATTRS
-_mm512_kunpackb (__mmask16 __A, __mmask16 __B)
-{
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS_CONSTEXPR
+_mm512_kunpackb(__mmask16 __A, __mmask16 __B) {
   return (__mmask16) __builtin_ia32_kunpckhi ((__mmask16) __A, (__mmask16) __B);
 }
 
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS_CONSTEXPR
+_mm512_kunpackd(__mmask64 __A, __mmask64 __B) {
+  return (__mmask64)__builtin_ia32_kunpckdi((__mmask64)__A, (__mmask64)__B);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS_CONSTEXPR
+_mm512_kunpackw(__mmask32 __A, __mmask32 __B) {
+  return (__mmask32)__builtin_ia32_kunpcksi((__mmask32)__A, (__mmask32)__B);
+}
+
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS_CONSTEXPR
 _mm512_kxnor(__mmask16 __A, __mmask16 __B) {
   return (__mmask16) __builtin_ia32_kxnorhi ((__mmask16) __A, (__mmask16) __B);

clang/test/CodeGen/X86/avx512f-builtins.c

@@ -9126,6 +9126,24 @@ __mmask16 test_mm512_kunpackb(__m512i __A, __m512i __B, __m512i __C, __m512i __D
                                                        __E, __F);
 }
 
+TEST_CONSTEXPR(_mm512_kunpackb(0xFF00, 0x00FF) == 0xFF00);
+TEST_CONSTEXPR(_mm512_kunpackb(0xABCD, 0x1234) == 0xCD34);
+TEST_CONSTEXPR(_mm512_kunpackb(0x00FF, 0xFF00) == 0x0000);
+TEST_CONSTEXPR(_mm512_kunpackb(0xAAAA, 0x5555) == 0xAA55);
+TEST_CONSTEXPR(_mm512_kunpackb(0x1234, 0xABCD) == 0x34CD);
+
+TEST_CONSTEXPR(_mm512_kunpackw(0xFFFF0000u, 0x0000FFFFu) == 0x0000FFFFu);
+TEST_CONSTEXPR(_mm512_kunpackw(0xABCD1234u, 0x56789ABCu) == 0x12349ABCu);
+TEST_CONSTEXPR(_mm512_kunpackw(0x0000FFFFu, 0xFFFF0000u) == 0x00000000u);
+TEST_CONSTEXPR(_mm512_kunpackw(0xAAAA5555u, 0x5555AAAAu) == 0x5555AAAAu);
+TEST_CONSTEXPR(_mm512_kunpackw(0x12345678u, 0xABCDEF12u) == 0x5678EF12u);
+
+TEST_CONSTEXPR(_mm512_kunpackd(0xFFFFFFFF00000000ull, 0x00000000FFFFFFFFull) == 0x00000000FFFFFFFFull);
+TEST_CONSTEXPR(_mm512_kunpackd(0xABCDEF0123456789ull, 0x0123456789ABCDEFull) == 0x234567899ABCDEFull);
+TEST_CONSTEXPR(_mm512_kunpackd(0x00000000FFFFFFFFull, 0xFFFFFFFF00000000ull) == 0x0000000000000000ull);
+TEST_CONSTEXPR(_mm512_kunpackd(0xAAAA5555AAAA5555ull, 0x5555AAAA5555AAAAull) == 0xAAAA55555555AAAAull);
+TEST_CONSTEXPR(_mm512_kunpackd(0x123456789ABCDEFull, 0xFEDCBA9876543210ull) == 0x89ABCDEF76543210ull);
+
 __mmask16 test_mm512_kxnor(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: test_mm512_kxnor
   // CHECK: [[LHS:%.*]] = bitcast i16 %{{.*}} to <16 x i1>